Review Reports

N° of comment

Comment

Answer

REVIEWER 1

1.1

Which is the additional contribution of this work when compared to one of the same authors presented in https://doi.org/10.1115/GT2023-104176 and https://doi.org/10.1115/GT2022-83550.

The abstract and introduction now better describe the contribution of the present paper relative to the previous ones.

1.2

The manuscript should include a comparison between the methods proposed and other methods existing in the literature, such as: 

• Lu, Feng, Jiang, Jipeng and Huang, Jinquan. "Gas Turbine Engine Gas-path Fault Diagnosis Based on Improved SBELM Architecture " International Journal of Turbo & Jet-Engines, vol. 35, no. 4, 2018, pp. 351-363. https://doi.org/10.1515/tjj-2016-0050
• Jingkai Zhang, Zhitao Wang, Shuying Li, Pengfei Wei,
  A digital twin approach for gas turbine performance based on deep multi-model fusion, Applied Thermal Engineering, Volume 246, 2024, 122954,
  ISSN 1359-4311, https://doi.org/10.1016/j.applthermaleng.2024.122954.

The introduction includes a qualitive comparison of the existing and proposed models. A quantitative comparison is difficult to realize because the models proposed have a new structure.

Because of a limited time for the paper revision, we were able to find only the first cited paper. It deals with a fault classification algorithm, but our paper is devoted to an engine model, and we cannot directly compare the engine model with the engine diagnostic algorithm. Instead of the comparison, we can state that the use of the proposed model will make much easier the input data generation for the algorithm. Actually, the data are generated by a thermodynamic model. Our models will reduce thousand time the data generation time. As shows our session 7, the diagnostic reliability of the algorithm will be practically the same.

1.3

Equation 1 appears without being introduced.

We added the sentence that introduces eq. (1)

1.4

In line 240, the authors refer to what concretely?

This is the description of the GPA approach. We slightly extended this paragraph and added a reference.

1.5

There are several problems with the formation of the text.

What type of problems, paper structure, English technical writing, or formatting?

1.6

Line 292 --> the values 0.05-0.07 are percentual? Which units apply if this is the case?

(5-7) % is implied. We made it clear in the text.

1.7

Lines 302-303. The sentence should be justified.

We added a reference.

1.8

The authors should clarify the matrix M in line 317.

We corrected the description of the matrix M and added the reference.

1.9

In lines 356-358, the authors argue about the calculation time. The authors should evidence some values and link those times to the computational resources used.

We added quantitative information about calculation time.

1.10

the authors should provide evidence of the consequences of the assumptions presented in lines 359-367

The consequences will not be considerable. We rewrote their discussion in the paragraph below the assumptions list.

1.11

the authors should justify the conclusion in lines 368-370.

The consequences will not be considerable. We rewrote their discussion in the paragraph below the assumptions list.

1.12

The authors refer to some SW, Matlab programs, etc., without evidence of where those programs may be found.

The programs for testing and comparing the proposed models were developed by us using Matlab and its Neural Network Toolbox. Two thermodynamic models are from Gas Turb and one model was developed by ourselves. All this information is mentioned in the paper.

1.13

the abstract and the conclusions should be rewritten to include some results.

Done

REVIEWER 2

2.1

In the abstract, the authors state that pattern recognition techniques (data classification) were used. However, the network presented in the paper is the MLP (Multilayer Perceptron), which is a data estimation or prediction network. It would be helpful to clarify the difference between these approaches in the text and how the MLP fits into the proposed analysis.

MLP can be not also an engine performance approximator, but also a good engine fault classifier. In this paper, we principally use the first MLP function to create an MLP-based engine model. We use the second function only in section 7, where MLP is a part of a diagnostic algorithm. These 2 MLP uses are quite different. The above MLP information is mentioned many times in sections 1 through 6 (first MLP function) and in section 7 (second function).

2.2

Additionally, how many samples were used for both training and validation of the network? Providing this information would help better understand the robustness of the model.

In input data sets, one engine operating point results in one input sample (pattern). Section 5 mentions these numbers: Engine 1 - 4595 operating points, Engine 2 - 7290 points, and Engine 3 - 40000 points. For all the engines, these data were randomly divided in the proportion 85% to 15% for learning and validation needs.

2.3

Given the data presented, which includes field measurements and test bench data, how did you ensure the generalization of the MLP (Multilayer Perceptron) model for different operating conditions? 

For different operating conditions, correction formulas should be used. This is described in the first assumption (see section 3, p 8).

2.4

Considering the accuracy variation observed in the test cases, what strategies were implemented to prevent overfitting and ensure the model maintains accuracy across a wide range of real-world scenarios?

To prevent overfitting, the EarlyStopping option was used. Additionally, we always checked the difference of approximation errors obtained on learning and validation data. This difference always was small (See Table 5). As to wide range conditions, the errors were verified for each monitored variable at each operating point by the corresponding plots (like those in Figures 4 and 5), and no outliers were found.

2.5

Furthermore, what limitations were encountered when applying the model to operating conditions outside of the training samples, and how were these addressed to ensure the model's applicability in online diagnostics?

Like any data-driven model, our models are accurate only in the area of training data. Extrapolation properties of the model were not studied. To ensure the applicability in real online diagnostics, all possible real operating ranges should be presented in training data.

2.6

Figure 4 shows significant differences between the desired (true) and estimated values. Could you explain the cause of this, and does it affect the validation results?

The objective of this paper is to present not only the results, but also the process of models’ creation. So, we do not hide possible difficulties. One of them is illustrated by Fig. 4. This accuracy problem is not general. It has arisen only in one estimated parameter of the inverse model of Engine 1. Subsection “Engine 1” in Section 6.2 and Section 8 explain the possible causes and probable solution of the problem.

REVIEWER 3

3.1

From the simulation point of view, the paper is too redundant a manuscript. The authors must reconsider their work and complete the article with a more concise and explicit interpretation.

We excluded some secondary information, primarily in the introduction. Despite some additional information asked by the reviewers and included in the total text, it has reduced from 946 to 870 lines.

3.2

State-of-the-art is not entirely related to the current concern of the nowadays solution and the research interest area. The authors must reconsider the introduction and must add at least one reference in extenso which treats a similar subject area, like fault diagnosis of gas engines.  From the introduction, the explanation of the authors (it is not about book chapter…) from row 161 to row 210, from my point of view should be deleted from the paper.

Nowadays solutions are mostly related to deep learning, but not diagnostic gas turbine models.

The comment “The authors must reconsider the introduction and must add at least one reference in extenso which treats a similar subject area, like fault diagnosis of gas engines” is not clear to us. The introduction refers to about 40 publications. All of them, except for 3 artificial intelligence books, are devoted to gas turbine diagnostics. And most of them are related to gas turbine diagnostic models.

The introduction was revised and reduced by 30%. The text in rows 161 to 210 was rewritten and now occupies 20 rows.

3.3

At chapter 2- is mostly adapted from a technical book and is not suitable to explain the mathematical modulation, is more interesting to present the novelty of the research. The authors must highlight their contribution and their innovative work in regard of presenting theoretical aspects presented from the theoretical background of the simulating tool used ( MatLab support), it is not bringing any value to the paper.

Chapter 2 describes existing models used in gas path diagnostics. We think it is not the place to show the novelty of the paper. Instead, the novelty is now highlighted in the abstract, introduction, Section 3, and conclusions.

3.4

Very hard to understand the explanation of the fig. 4 and fig. 5… what are the inputs for the true values? In what scenario were obtained the estimated values?

Fig. 4 shows erroneous estimation (parameter δΘ₁ of fan capacity) and Figure 5 - correct estimation of all other component performances. The true values of fault parameters are those introduced in the thermodynamic model of Engine 1 during data generation, and estimated values are results of the inverse model. Some corrections were made to make it more clear.

3.5

At discussions must be reconsidered and made more explicit, the abstract interpretation does not offer any proof of expertise from the authors

Discussion has concrete detailed information about the accuracy and applicability of the developed models. We do not know how to make the discussion more explicit.

We included in abstract some information about our expertise.

3.6

Engines 2 and 3 are more accurate…. But why are more accurate?

Engines 2 and 3 have well-prepared input data. Engine 1 has some inconsistency of input data. This problem is discussed in detail in the subsection “Engine 1” of section 6.2.

3.7

The conclusions are too comprehensive, in contrast with the first 2 chapters are too much redundant and detailed ( not that much necessary, these can be compressed using better references). In Conclusions the authors must highlight the reasons and the self-contributions to the results obtained  

What conclusion do you imply? The paper conclusions occupy only two small paragraphs. And it is not a proper place for references to previous works. We added some numbers to the paper conclusions.

3.8

The paper can be useful for the research area and can be published in the journal, but needs major adjustments.

We tried to understand and realize all the comments.